File size: 2,106 Bytes
62d0c8c 75cb516 62d0c8c 75cb516 62d0c8c 75cb516 62d0c8c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 | import streamlit as st
def run():
st.header("🧪 Methodology")
st.write("How we turned ~5,400 Tokopedia reviews into actionable themes.")
st.markdown("#### The Pipeline")
st.markdown("`Reviews -> Clean text -> Embeddings -> Cluster (BERTopic) -> Themes`")
st.markdown(
"""
1. **Data** - PRDECT-ID: ~5,400 Indonesian product reviews, each labeled Positive/Negative plus an emotion.
2. **Preprocessing** - lowercase, strip URLs/symbols, normalize slang (`gak` -> `tidak`); Indonesian
stopwords are removed only when extracting the topic keywords.
3. **Embeddings** - each review becomes a 384-dim vector using a multilingual language model
(paraphrase-multilingual-MiniLM-L12-v2).
4. **Clustering** - reviews are split by sentiment, then BERTopic (UMAP + HDBSCAN) groups similar
vectors into topics, separately for positive and negative reviews.
5. **Themes** - the many micro-topics are grouped into **8 praise** and **9 complaint** business themes.
"""
)
st.markdown("#### Why BERTopic?")
c1, c2, c3 = st.columns(3)
c1.markdown("**K-Means**\n\nSimple, but needs the number of clusters up front and gave coarse, "
"mixed groups.")
c2.markdown("**LDA**\n\nClassic bag-of-words baseline; broad topics that ignore word context.")
c3.markdown("**BERTopic (chosen)**\n\nEmbeddings + density clustering. Best coherence and "
"diversity, and the finest themes.")
st.markdown("#### The Key Decision")
st.info(
"BERTopic's default clustering (HDBSCAN 'eom') collapsed **89% of complaints into one "
"cluster**. Switching to 'leaf' separated them into 9 clean themes - our single most "
"important tuning choice."
)
st.markdown("#### How This App Predicts")
st.markdown(
"Your review goes through the **same cleaning**, becomes an embedding, and is matched by "
"**cosine similarity** to the learned topic vectors. The nearest theme - and whether it "
"reads as praise or a complaint - is returned, with no retraining needed."
)
|